You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[MIGraphX EP] Add migx ep fp8 support and int4 weights (microsoft#23534)
* Add fp8 and int4 types in supported list for Onnxruntime EP
* Add support for int4 inputs
Map things to int8 right now as we don't explicitly set an int4 input
type and pack/unpack int4 operands
* Add flag to allow for fp8 quantization through Onnxruntime API
* Add fp8 quantization to the compile stage of the MIGraphX EP
Mirror the same calibration code we use for int8 and just change which
quantize we call through the MIGraphx API
* cleanup logging
* Cleanup and encapsulate quantization / compile functions
- Add additional flags for fp8 thats shared for int8
- Add lockout warning message when int8/fp8 used at the same time
* Run lintrunner pass
* Fix session options inputs + add better logging.
Previous runs using session options failed as we were missing pulling in
inputs from the python interface. This plus additional logging allowed
me to track what options were invoked via env and what were added during
the start of an inference session
* Fix naming for save/load path varibles to be consistent with enable.
* Print only env variables that are set as warnings
need this so the user knows there's any of the environment variables
running in the background to ensure proper consistently between runs.
---------
### Description
<!-- Describe your changes. -->
Changes to cleanup the MIGraphX EP quantization code as well as adding
fp8 quantization support along with int4 support.
Cleanup changes handle a few instances of issues seen with the python
interface when taking in provider options
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Required as we fix ignored flags when using provider_options for the
MIGraphX EP
Adding fp8 quantization through the MIGraphX API
Adding int4 weight support for packed int4 weights for MIGraphX
inference
---------
Co-authored-by: Ted Themistokleous <[email protected]>
0 commit comments